177 Clinical sub-phenotypes of Staphylococcus aureus bacteraemia 8 method for dealing with missing data in LCA20. With FIML, data is not imputed, but all available information is used for calculation of the likelihood contribution of each respondent to the estimation of the model parameters20. Model selection was based on a combination of statistical criteria and clinical knowledge. The statistical criteria used were the Bayesian Information Criteria (BIC), number of classes, and size of smallest class. The BIC is a statistical measure that provides information on the model fit, and is best at identifying the correct number of classes if a combination of continuous and categorical data is used20. A decrease in the BIC suggests that the addition of more classes is worth the added model complexity21. To avoid a local maximum, in which case it would be difficult to replicate our findings, 16 random starting values were used, and 50 iterations for each start value. Those solutions were checked to make sure that the same maximum likelihood solution was found. When setting the seed, a fixed starting point for random number generation was established, which ensures reproducibility across different runs of the analysis. After identification of classes, we estimated the posterior probability of class membership for each of the identified classes for each individual, and assigned the individual to the class with the highest probability20. Given that LCA is a probabilistic method, there is a certain degree of uncertainty in class assignment, which can lead to classification errors. For example, an individual may have a 0·9 chance of belonging to class one, and a 0·1 chance of belonging to class two. This individual is then assigned to class one. We correct for misclassification error using the biasadjusted three-step LCA22. LCA was done using the Latent GOLD 6.0 statistical software package23. Cohort characteristics were compared using contingency tables for categorical variables (Chi-square or Fisher’s exact test), and Mann Whitney or Kruskal-Wallis tests for continuous variables (which Shapiro-Wilk tests demonstrated to be not normally distributed). To compare class-defining variables between sub-phenotypes, z-scores were calculated (z=((value for subphenotype -mean for variable))/(standard deviation for variable)). Additional meta-data not included as class-defining variables was compared between patients stratified by predicted sub-phenotype membership. Unadjusted one-year survival was compared using a Kaplan-Meier survival curve and log-rank test, performed using the survminer24 and ggplot225 packages in R (RStudio version Version 2023·06·1+524). Unless otherwise stated, analyses and data visualisation were done using GraphPad Prism Version 10·0·3 for macOS. Results Cohort characteristics Characteristics of the Edinburgh, ARREST and SAFO cohorts are compared in Table 1. In comparison with the Edinburgh cohort, patients in ARREST were more likely to have SAB originating from skin or soft tissue infection (SSTI), and patients in SAFO were more likely to have an intravenous catheter as the source of bacteraemia.
RkJQdWJsaXNoZXIy MTk4NDMw